一、背景
0、Hybrid Transformer 论文解读
1、代码复现|Demucs Music Source Separation_demucs架构原理-CSDN博客
2、Hybrid Transformer 各个模块对应的代码具体在工程的哪个地方
3、Hybrid Transformer 各个模块的底层到底是个啥（初步感受）？
4、Hybrid Transformer 各个模块处理后，数据的维度大小是咋变换的？
5、Hybrid Transformer 拆解STFT模块
6、Hybrid Transformer 拆解频域编码模块
从模块上划分，Hybrid Transformer Demucs 共包含 （STFT模块、时域编码模块、频域编码模块、Cross-Domain Transformer Encoder模块、时域解码模块、频域解码模块、ISTFT模块）7个模块。已完成解读：STFT模块、频域编码模块（时域编码和频域编码类似，后续不再解读时域编码模块），待解读：Cross-Domain Transformer Encoder模块。
本篇目标：拆解频域解码模块、ISTFT模块的底层。时域解码和频域解码原理类似（后续不再拆解时域解码模块）。
二、频域解码模块
class HDecLayer(nn.Module):def __init__(self, chin, chout, last=False, kernel_size=8, stride=4, norm_groups=1, empty=False,freq=True, dconv=True, norm=True, context=1, dconv_kw={}, pad=True,context_freq=True, rewrite=True):"""Same as HEncLayer but for decoder. See `HEncLayer` for documentation."""super().__init__()norm_fn = lambda d: nn.Identity() # noqaif norm:norm_fn = lambda d: nn.GroupNorm(norm_groups, d) # noqaif pad:pad = kernel_size // 4else:pad = 0self.pad = padself.last = lastself.freq = freqself.chin = chinself.empty = emptyself.stride = strideself.kernel_size = kernel_sizeself.norm = normself.context_freq = context_freqklass = nn.Conv1dklass_tr = nn.ConvTranspose1dif freq:kernel_size = [kernel_size, 1]stride = [stride, 1]klass = nn.Conv2dklass_tr = nn.ConvTranspose2dself.conv_tr = klass_tr(chin, chout, kernel_size, stride)self.norm2 = norm_fn(chout)if self.empty:returnself.rewrite = Noneif rewrite:if context_freq:self.rewrite = klass(chin, 2 * chin, 1 + 2 * context, 1, context)else:self.rewrite = klass(chin, 2 * chin, [1, 1 + 2 * context], 1,[0, context])self.norm1 = norm_fn(2 * chin)self.dconv = Noneif dconv:self.dconv = DConv(chin, **dconv_kw)def forward(self, x, skip, length):if self.freq and x.dim() == 3:B, C, T = x.shapex = x.view(B, self.chin, -1, T)if not self.empty:x = x + skipif self.rewrite:y = F.glu(self.norm1(self.rewrite(x)), dim=1)else:y = xif self.dconv:if self.freq:B, C, Fr, T = y.shapey = y.permute(0, 2, 1, 3).reshape(-1, C, T)y = self.dconv(y)if self.freq:y = y.view(B, Fr, C, T).permute(0, 2, 1, 3)else:y = xassert skip is Nonez = self.norm2(self.conv_tr(y))print('self.pad,self.last:', self.pad,self.last)if self.freq:if self.pad:z = z[..., self.pad:-self.pad, :]else:z = z[..., self.pad:self.pad + length]assert z.shape[-1] == length, (z.shape[-1], length)if not self.last:z = F.gelu(z)return z, y
频域解码模块的核心代码如上所示。在上一篇频域编码模块的基础上，继续贴出完善之后的频域编解码模块全景图。
编码层:Conv2d+Norm1+GELU, Norm1：Identity()
解码层:(Conv2d+Norm1+GLU)+(ConvTranspose2d+Norm2+倒数第二个维度裁剪+GELU), Norm1\Norm2：Identity()
残差连接：（Conv1d+GroupNorm+GELU +Conv1d+GroupNorm+GLU+LayerScale()）+（Conv2d+Norm2+GLU），Norm2：Identity() ，备注：Identity可以理解成直通
#频域编码层1-4的Conv2d分别是：
Conv2d(4, 48, kernel_size=(8, 1), stride=(4, 1), padding=(2, 0))
Conv2d(48, 96, kernel_size=(8, 1), stride=(4, 1), padding=(2, 0))
Conv2d(96, 192, kernel_size=(8, 1), stride=(4, 1), padding=(2, 0))
Conv2d(192, 384, kernel_size=(8, 1), stride=(4, 1), padding=(2, 0))#频域解码层4-1的Conv2d和ConvTranspose2d
Conv2d(384, 768, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ConvTranspose2d(384, 192, kernel_size=(8, 1), stride=(4, 1))
Conv2d(192, 384, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ConvTranspose2d(192, 96, kernel_size=(8, 1), stride=(4, 1))
Conv2d(96, 192, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ConvTranspose2d(96, 48, kernel_size=(8, 1), stride=(4, 1))
Conv2d(48, 96, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ConvTranspose2d(48, 16, kernel_size=(8, 1), stride=(4, 1))
残差连接模块如下所示。
#残差连接1
DConv((layers): ModuleList((0): Sequential((0): Conv1d(48, 6, kernel_size=(3,), stride=(1,), padding=(1,))(1): GroupNorm(1, 6, eps=1e-05, affine=True)(2): GELU(approximate=none)(3): Conv1d(6, 96, kernel_size=(1,), stride=(1,))(4): GroupNorm(1, 96, eps=1e-05, affine=True)(5): GLU(dim=1)(6): LayerScale())(1): Sequential((0): Conv1d(48, 6, kernel_size=(3,), stride=(1,), padding=(2,), dilation=(2,))(1): GroupNorm(1, 6, eps=1e-05, affine=True)(2): GELU(approximate=none)(3): Conv1d(6, 96, kernel_size=(1,), stride=(1,))(4): GroupNorm(1, 96, eps=1e-05, affine=True)(5): GLU(dim=1)(6): LayerScale()))
)
Conv2d(48, 96, kernel_size=(1, 1), stride=(1, 1))#残差连接2
DConv((layers): ModuleList((0): Sequential((0): Conv1d(96, 12, kernel_size=(3,), stride=(1,), padding=(1,))(1): GroupNorm(1, 12, eps=1e-05, affine=True)(2): GELU(approximate=none)(3): Conv1d(12, 192, kernel_size=(1,), stride=(1,))(4): GroupNorm(1, 192, eps=1e-05, affine=True)(5): GLU(dim=1)(6): LayerScale())(1): Sequential((0): Conv1d(96, 12, kernel_size=(3,), stride=(1,), padding=(2,), dilation=(2,))(1): GroupNorm(1, 12, eps=1e-05, affine=True)(2): GELU(approximate=none)(3): Conv1d(12, 192, kernel_size=(1,), stride=(1,))(4): GroupNorm(1, 192, eps=1e-05, affine=True)(5): GLU(dim=1)(6): LayerScale()))
)
Conv2d(96, 192, kernel_size=(1, 1), stride=(1, 1))#残差连接3
DConv((layers): ModuleList((0): Sequential((0): Conv1d(192, 24, kernel_size=(3,), stride=(1,), padding=(1,))(1): GroupNorm(1, 24, eps=1e-05, affine=True)(2): GELU(approximate=none)(3): Conv1d(24, 384, kernel_size=(1,), stride=(1,))(4): GroupNorm(1, 384, eps=1e-05, affine=True)(5): GLU(dim=1)(6): LayerScale())(1): Sequential((0): Conv1d(192, 24, kernel_size=(3,), stride=(1,), padding=(2,), dilation=(2,))(1): GroupNorm(1, 24, eps=1e-05, affine=True)(2): GELU(approximate=none)(3): Conv1d(24, 384, kernel_size=(1,), stride=(1,))(4): GroupNorm(1, 384, eps=1e-05, affine=True)(5): GLU(dim=1)(6): LayerScale()))
)
Conv2d(192, 384, kernel_size=(1, 1), stride=(1, 1))#残差连接4
DConv((layers): ModuleList((0): Sequential((0): Conv1d(384, 48, kernel_size=(3,), stride=(1,), padding=(1,))(1): GroupNorm(1, 48, eps=1e-05, affine=True)(2): GELU(approximate=none)(3): Conv1d(48, 768, kernel_size=(1,), stride=(1,))(4): GroupNorm(1, 768, eps=1e-05, affine=True)(5): GLU(dim=1)(6): LayerScale())(1): Sequential((0): Conv1d(384, 48, kernel_size=(3,), stride=(1,), padding=(2,), dilation=(2,))(1): GroupNorm(1, 48, eps=1e-05, affine=True)(2): GELU(approximate=none)(3): Conv1d(48, 768, kernel_size=(1,), stride=(1,))(4): GroupNorm(1, 768, eps=1e-05, affine=True)(5): GLU(dim=1)(6): LayerScale()))
)
Conv2d(384, 768, kernel_size=(1, 1), stride=(1, 1))
三、ISTFT模块
ISTFT模块的核心代码如下所示。
import torch as th
def ispectro(z, hop_length=None, length=None, pad=0):*other, freqs, frames = z.shapen_fft = 2 * freqs - 2z = z.view(-1, freqs, frames)win_length = n_fft // (1 + pad)is_mps = z.device.type == 'mps'if is_mps:z = z.cpu()x = th.istft(z,n_fft,hop_length,window=th.hann_window(win_length).to(z.real),win_length=win_length,normalized=True,length=length,center=True)_, length = x.shapereturn x.view(*other, length)其中，torch.istft【逆短时傅里叶变换（Inverse Short Time Fourier Transform，ISTFT）】，该函数期望是torch.stft函数的逆过程。它具有相同的参数（加上一个可选参数length），并且应该返回原始信号的最小二乘估计。算法将根据NOLA条件（非零重叠）进行检查。
#### torch.istft接口参数####
input (Tensor): 输入张量，期望是`torch.stft`的输出，可以是复数形式（`channel`, `fft_size`, `n_frame`），或者是实数形式（`channel`, `fft_size`, `n_frame`, 2），其中`channel`维度是可选的。deprecated:: 1.8.0
实数输入已废弃，请使用`stft(..., return_complex=True)`返回的复数输入代替。
n_fft (int): 傅里叶变换的大小。
hop_length (Optional[int]): 相邻滑动窗口帧之间的距离。（默认：`n_fft // 4`）
win_length (Optional[int]): 窗口帧和STFT滤波器的大小。（默认：`n_fft`）
window (Optional[torch.Tensor]): 可选的窗函数。（默认：`torch.ones(win_length)`）
center (bool): 指示输入是否在两边进行了填充，使得第`t`帧位于时间`t × hop_length`处居中。（默认：`True`）
normalized (bool): 指示STFT是否被标准化。（默认：`False`）
onesided (Optional[bool]): 指示STFT是否为单边谱。（默认：如果输入尺寸中的`n_fft != fft_size`则为`True`）
length (Optional[int]): 修剪信号的长度，即原始信号的长度。（默认：整个信号）
return_complex (Optional[bool])：指示输出是否应为复数，或者输入是否应假定源自实信号和窗函数。注意，这与`onesided=True`不兼容。（默认：`False`）
频域解码模块和ISTFT模块解读完毕。还剩一个Cross-Domain Transformer Encoder模块没有解读。后面又来新的活了，希望能把demucs落地~。
感谢阅读，最近开始写公众号（分享好用的AI工具），欢迎大家一起见证我的成长（桂圆学AI）